You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
but forgot that v had a pos dimension not a key_pos dimension (it should have had a key_pos but they forgot to rename. attn had both, so this produced a result with the right shape, but it didn't do what they wanted.
This is a valid operation and it's not inconceivable someone could want it: it's like einsum("ij,je->je, attn, v) or hax.sum(attn, "key_pos") * v. However it seems like an uncommon thing? should we maybe issue a warning? or have a "paranoid debug" mode that logs warnings for things like this?
The text was updated successfully, but these errors were encountered:
dlwh
changed the title
Maybe warn if
Maybe warn if dot reduces a dim that isn't in common between args?
Aug 24, 2023
Someone did:
but forgot that
v
had apos
dimension not akey_pos
dimension (it should have had akey_pos
but they forgot to rename.attn
had both, so this produced a result with the right shape, but it didn't do what they wanted.This is a valid operation and it's not inconceivable someone could want it: it's like
einsum("ij,je->je, attn, v)
orhax.sum(attn, "key_pos") * v
. However it seems like an uncommon thing? should we maybe issue a warning? or have a "paranoid debug" mode that logs warnings for things like this?The text was updated successfully, but these errors were encountered: